burakkizil/Auteur-Dataset

Auteur Dataset

arXiv | Webpage | Trajectory Generation Demo

This dataset accompanies the paper Auteur. It provides instruction-tuning data designed to train large language models in generating synchronized, human-centric actor and camera trajectories for controllable video generation.

Dataset Overview

The dataset is provided as a JSON file formatted as conversational exchanges between a user (human) and the model (gpt). It trains the model to sequentially plan an actor's movement in the scene and then plan the camera's framing response relative to that actor.

Structure

Each conversational sequence strictly follows this two-step autoregressive generation process:

• Actor Motion (Human Trajectory): A natural language prompt describes the subject's movement. The model outputs a JSON array describing the coarse actor's trajectory (translation and yaw).
• Camera Motion (DSL): A prompt providing a natural language camera description alongside the initial camera state (orientation, depth, framing, etc.). The model outputs a JSON object containing the updated cinematography Domain-Specific Language (DSL) state and modifier tags (e.g., static, jitter, ease).

Example Entry

[
  {
    "from": "human",
    "value": "The man, wearing a headscarf and a neutral expression, remains completely still throughout the entire shot. Generate Human Trajectory:"
  },
  {
    "from": "gpt",
    "value": "[{\"movements\": [{\"tx\": \"no\", \"ty\": \"no\", \"tz\": \"no\", \"yaw\": \"0\"}, {\"tx\": \"no\", \"ty\": \"no\", \"tz\": \"no\", \"yaw\": \"0\"}]}]"
  },
  {
    "from": "human",
    "value": "Based on this caption The camera maintains a static position at coordinates (1.0, 2.0, 3.0), capturing the subject in a dimly lit, mysterious atmosphere with soft, ambient lighting. A zoom-in (focal length change from 50mm to 100mm) transitions the framing from a medium shot to a close-up., continue this DSL {\"orientation\": \"front\", \"depth\": \"medium_close_up\", \"framing\": \"center\", \"look_at_level\": \"eye_level\", \"camera_level\": \"eye_level\", \"roll\": \"normal\"}:"
  },
  {
    "from": "gpt",
    "value": "{\"static\": \"yes\", \"jitter\": \"no\", \"ease\": \"no\", \"direction\": \"none\", \"DSL\": {\"orientation\": \"front\", \"depth\": \"medium_close_up\", \"framing\": \"center\", \"look_at_level\": \"eye_level\", \"camera_level\": \"eye_level\", \"roll\": \"normal\"}}"
  }
]

Camera Parameters Conversion

The model outputs an actor-relative cinematography DSL. To obtain explicit 6-DoF world-space camera parameters, you must convert this DSL using the generated actor trajectory as an anchor.

Please refer to the Demo & Code for the conversion scripts, or consult the mathematical formulations provided in our paper.

Citation

If you use this dataset, please cite the following paper:

@misc{kizil2026auteur,
      title={Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation}, 
      author={Muhammed Burak Kizil and Enes Sanli and Niloy J. Mitra and Xuelin Chen and Erkut Erdem and Aykut Erdem and Duygu Ceylan},
      year={2026},
      eprint={2606.01900},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.01900}, 
}